ESTRO 2021 Abstract Book

S366

ESTRO 2021

Figure 2. MAEs and pSNR calculated within the body, soft tissue, and bone with respect to the U-net and proposed model. Conclusion Qualitative and quantitative evaluation indicates that the proposed method can synthesize sCT with accurate CT numbers and best texture information, especially in soft tissue and the body. The achievement can increase treatment precision with sCT and show the potentiality in sCT-based organs contouring and dose calculation for adaptive radiotherapy. OC-0478 Neural network based synthetic CTs for adaptive proton therapy of lung cancer A. Thummerer 1 , P. Zaffino 2 , C. Seller Oria 1 , A. Meijers 1 , G. Guterres Marmitt 1 , J. Seco 3,4 , J. Langendijk 1 , A. Knopf 1,5 , M.F. Spadea 2 , S. Both 1 1 University Medical Center Groningen, Department of Radiation Oncology, Groningen, The Netherlands; 2 Magna Graecia University Catanzaro, Department of Experimental and Clinical Medicine, Catanzaro, Italy; 3 Deutsches Krebsforschungszentrum (DKFZ), Department of Biomedical Physics in Radiation Oncology, Heidelberg, Germany; 4 Heidelberg University, Department of Physics and Astronomy, Heidelberg, Germany; 5 University Hospital of Cologne, Department I of Internal Medicine, Center for Integrated Oncology Cologne, Cologne, Germany Purpose or Objective Adaptive proton therapy (APT) accounts for anatomical and physiological changes to ensure target coverage and organ-at-risk sparing during the entire treatment course. An imaging modality that may detect these anatomical changes is cone-beam computed tomography (CBCT). However, CBCT-images suffer from severe image artifacts (e.g. scatter) that hinder accurate proton dose calculations. Deep convolutional neural networks (DCNN) have shown potential to correct CBCTs and create synthetic CTs (sCTs) that enable proton dose calculations in various anatomical locations (e.g head&neck, pelvis). In this study such a DCNN together with an accompanying planning CT (pCT) based patient specific correction technique was used to generate sCTs and their suitability for adaptive proton therapy of lung cancer patients was evaluated in terms of image quality and dosimetric accuracy. Materials and Methods A dataset consisting of CBCT- and same-day repeat CT-images from 33 lung cancer patients, treated with proton therapy, was used to train and evaluate the DCNN. 3-fold cross validation was employed to utilize all 33 patients for image and dosimetric evaluation. After the DCNN-conversion, an automatic patient specific correction method, using a smoothed and truncated difference map between the pCT and sCT, was introduced, mainly to correct CT-numbers of lung tissue, which are difficult to generate consistently and accurately by the DCNN. For image quality assessment, mean absolute error (MAE) and mean error (ME) were calculated for sCTs with (sCT cor ) and without (sCT orig ) the pCT-based correction method. For the dosimetric evaluation, clinical treatment plans were recalculated on both synthetic CTs and gamma pass ratios (3%/3mm) were used to compare dose distributions to those calculated on the reference CT scans. Results Figure 1 shows an overview of CBCT, CT, sCT orig and sCT cor for patient 5 together with difference images between sCTs and CT. Average MAEs (ME) of 34.7±7.2 HU (5.2±9.8 HU) and 30.8±4.7 (2.7±4.8 HU) were observed for sCT orig and sCT cor respectively. The recalculation of clinical treatment plans resulted in average gamma pass ratios of 93.9±4.6 % for sCT orig and 96.9±2.2 % for sCT cor . Results from the gamma analysis are presented in Figure 2 for each patient individually.